Large language models (LLMs) are capable of generating plausible explanations of how they arrived at an answer to a question. However, these explanations can misrepresent the model's "reasoning" process, i.e., they can be unfaithful. This, in turn, can lead to over-trust and misuse. We introduce a new approach for measuring the faithfulness of LLM explanations. First, we provide a rigorous definition of faithfulness. Since LLM explanations mimic human explanations, they often reference high-level concepts in the input question that purportedly influenced the model. We define faithfulness in terms of the difference between the set of concepts that LLM explanations imply are influential and the set that truly are. Second, we present a novel method for estimating faithfulness that is based on: (1) using an auxiliary LLM to modify the values of concepts within model inputs to create realistic counterfactuals, and (2) using a Bayesian hierarchical model to quantify the causal effects of concepts at both the example- and dataset-level. Our experiments show that our method can be used to quantify and discover interpretable patterns of unfaithfulness. On a social bias task, we uncover cases where LLM explanations hide the influence of social bias. On a medical question answering task, we uncover cases where LLM explanations provide misleading claims about which pieces of evidence influenced the model's decisions.

The integration of large language models (LLMs) on low-power edge devices such as Raspberry Pi, known as edge language models (ELMs), has introduced opportunities for more personalized, secure, and low-latency language intelligence that is accessible to all. However, the resource constraints inherent in edge devices and the lack of robust ethical safeguards in language models raise significant concerns about fairness, accountability, and transparency in model output generation. This paper conducts a comparative analysis of text-based bias across language model deployments on edge, cloud, and desktop environments, aiming to evaluate how deployment settings influence model fairness. Specifically, we examined an optimized Llama-2 model running on a Raspberry Pi 4; GPT 4o-mini, Gemini-1.5-flash, and Grok-beta models running on cloud servers; and Gemma2 and Mistral models running on a MacOS desktop machine. Our results demonstrate that Llama-2 running on Raspberry Pi 4 is 43.23% and 21.89% more prone to showing bias over time compared to models running on the desktop and cloud-based environments. We also propose the implementation of a feedback loop, a mechanism that iteratively adjusts model behavior based on previous outputs, where predefined constraint weights are applied layer-by-layer during inference, allowing the model to correct bias patterns, resulting in 79.28% reduction in model bias.

The city of Detroit has agreed to pay 300,000 to a Black man who was wrongly arrested for shoplifting and to also change how police use facial recognition technology to solve crimes after the software identified him as a suspect. The conditions are part of a lawsuit settlement with Robert Williams. His driver's license photo was incorrectly flagged by facial recognition software as a likely match to a man seen on security video at a Shinola watch store in 2018. "We are extremely excited that going forward there will be more safeguards on the use of this technology with our hope being to live in a better world because of it," Williams told reporters, "even though what we would like for them to do is not use it at all." The agreement was announced Friday by the American Civil Liberties Union and the civil rights litigation initiative at University of Michigan Law School.

Generative adversarial networks (GANs) have seen widespread adoption in machine learning, especially in computer vision applications. These "generative" models are capable of producing artificial images in many instances indistinguishable from the real thing. The most common use of these networks is that of artificial face generation. These so-called "deepfakes" have been used in a number of research and commercial applications. With their proliferation, however, have come predictable problems of bias in their generation. All such models are trained on large datasets. Several pre-trained models for StyleGANs 2 and 3 are trained on the Flickr (FFHQ) dataset.

When it first opened in theatres, in the fall of 1984, "Stop Making Sense," directed by Jonathan Demme and starring the rock group Talking Heads, was quickly recognized as one of the finest concert films ever made. Reviewer after reviewer settled on the word "exhilarating" to describe the experience of watching an expanded nine-member iteration of the four-piece group perform sixteen of their best-known songs in an uninterrupted sequence of dynamically staged and photographed musical vignettes. In the pages of this magazine, Pauline Kael praised the film as "close to perfection," and described the Heads front man, David Byrne, as "a stupefying performer." "He's so white he's almost mock-white," Kael wrote, "and so are his jerky, long-necked, mechanical-man movements. He seems fleshless, bloodless; he might almost be a Black man's parody of how a clean-cut white man moves. But Byrne himself is the parodist, and he commands the stage by his hollow-eyed, frosty verve."

In my work as an executive and the work I do today, it's somewhat taken at face value that I am a Black man,

All faces are not considered equal by Artificial Intelligence (AI) systems! A typical commercial AI face recognition system most accurately predicts fair-skinned males. The accuracy for detecting dark-skinned women is lower by over 30 percentage points, says Joy Buolamwini, Founder, Algorithmic Justice League. AI computer vision systems will appropriately classify and label the image of a bride in a typical western gown. However, when it was asked to classify the image of an Indian bride wearing a red sari, it is classified as an event, a costume or performing art.

While 2021 was an exciting year for AI regarding innovations and new inventions, it was immune to controversies and scandals. In this article, we take a look at some of the most prominent ones that grabbed headlines. From the range of announcements made at the Tesla AI Day 2021, one that caught the fancy of a lot of people was the humanoid robot. Introduced in a unique manner, a human dressed in a white bodysuit and shiny mask did the news reveal during the event. Called Optimus, this humanoid robot, standing five feet eight inches and weighing 125 pounds, would be capable of performing repetitive tasks; the first prototype is likely to be released next year.

Care and Feeding is Slate's parenting advice column. Have a question for Care and Feeding? Submit it here or post it in the Slate Parenting Facebook group. My nephew Nicolas is 9 and only plays video games. He wakes up around 2:00 pm doesn't brush his teeth or shower and begins playing games.

Those that haven't, are struggling to survive. We aren't, and what sets us apart, keeps us alive, and on top, is our intelligence. Artificial Intelligence, or AI, allows us to clone this great gift, i.e. human intelligence, using machines. In this article, we won't cover what's widely known -- how AI is changing the world for the better, from suggesting what you should watch next on Netflix to detecting signs of cancer in medical images better than doctors. We will instead focus on what's not often talked about -- how human bias can creep into AI, and how to avoid it.